Abstract
Snoring is the earliest and the most common symptom of Obstructive Sleep Apnea (OSA) . Quantitative analysis of snoring, however, is not used at present in the clinical diagnosis of the disease. Several researchers have reported differences in the formant frequencies of Apneic and benign snoring sounds (SS) based on linear prediction coding (LPC) analysis. However, SS is complex signal and at local low signal to noise ratio (SNR) . This signal complexity should reduce the accuracy of formant estimation. In this paper, we propose a novel approach to the diagnosis of OSA based on the formants of SSs, extracted via a noise-robust linear prediction technique. The proposed method and existing LPC-based method are compared via a measure, σ which indicates the standard deviation of first formant frequencies. The performance of the proposed method was evaluated on a database of clinical snoring sounds recorded overnight in the laboratory of a hospital sleep diagnostic center. Compared with existing LPC-based method, we show that the proposed method can differentiate (sensitivity: 88.9%, specificity: 88.9%, AUC: 0.85) between benign snoring (Apnea Hypopnea Index, AHI=6.0±3.2 event/h; 6188 episodes) and apneic snoring (AHI=40.7±20.2 event/h; 14066 episodes) .
